KorDef-LLM: Korean Defense Domain Instruction Corpus and Source-Grounded Evaluation Set

# KorDef-LLM: Korean Defense Domain Instruction Corpus and Evaluation Set This Zenodo deposit contains the openly released artifacts accompanying the paper: > **An Open Pipeline for Domain-Adaptive Instruction Tuning of Korean Defense Large Language Models** > Sang-Hwan Gwak, Ji-Young Choi, Chang-Hoo Jeong, Gunwoo Lee, Ina Kim, Kyung-Ha Lee > Korea Institute of Science and Technology Information (KISTI), 2026 ## Contents | File | Size | Description | |---|---|---| | `kordef_prompt_generated_clean.jsonl` | 127 MB | Recovered Korean defense instruction--response corpus (235,367 cleaned QA pairs) | | `eval_qa_329_final.jsonl` | 1.2 MB | Source-grounded evaluation set (329 candidates; 323 used in the paper after alignment audit) | | `cleaned_segments_all.jsonl` | 123 MB | 34,472 source-aware text segments extracted from public Korean Ministry of National Defense administrative-rule PDFs | | `source_manifest_all.csv` | 548 KB | Manifest of 2,540 public source PDFs with document IDs, titles, paths, and metadata | | `kordef_train_eval_overlap_report.txt` | 471 B | Train/evaluation overlap audit (exact question: 0; exact answer: 0; near-question Jaccard ≥ 0.80: 0) | | `kordef_prompt_generated_removed_artifacts.jsonl` | 49 KB | 58 Cyrillic/mixed-script artifact rows removed during cleaning | | `SHA256SUMS.txt` | 903 B | SHA256 hashes for integrity verification | ## File Schemas ### kordef_prompt_generated_clean.jsonl ### eval_qa_329_final.jsonl The paper uses 323 of these items after a final alignment audit removed 6 items with noisy table or form-like contexts. ## Train/Evaluation Independence | Audit | Count | |---|---| | Exact question overlaps | 0 | | Exact answer overlaps | 0 | | Near-question overlaps (Jaccard ≥ 0.80) | 0 | ## Data Provenance and Ethics All artifacts in this deposit are derived exclusively from public, unclassified Korean defense administrative rules and educational materials. No classified documents, operational logs, personal data, or restricted internal military records are included. The associated KorDef-LLM model is not intended for autonomous operational use and should not be deployed for military decision-making, targeting, procurement, maintenance, or safety-critical procedures without retrieval grounding, institutional security review, and human expert oversight. ## Related Resources - **Trained model weights**: https://huggingface.co/jeong0313/koni-it-m-7875 - **Code and evaluation scripts**: [TODO: GitHub URL once published] - **Paper**: An Open Pipeline for Domain-Adaptive Instruction Tuning of Korean Defense Large Language Models. PeerJ Computer Science, 2026 (under review). ## License The data are released under the **Creative Commons Attribution 4.0 International (CC-BY-4.0)** license. See `LICENSE.txt`. ## Citation If you use these artifacts, please cite the accompanying paper and this Zenodo deposit. See `CITATION.cff` for machine-readable citation metadata. ## Funding This work was supported by the Future Defense Bridge Technology Development Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) and the Defense Acquisition Program Administration (DAPA) of the Korea government under Grant RS-2024-00452972. ## Contact Corresponding author: Kyung-Ha Lee ([e-mail]) Large-Scale AI Research Center, KISTI, Daejeon, Republic of Korea